Abstract

Comminution is the most energy-intensive stage in mineral processing, representing up to 50% of the total energy requirements of mining complexes. Specialized particle size reduction and classification equipment are arranged in circuits to make the compounded effects work towards the required reduced particle size distribution for a given feed volume. Although in this context high energy consumption levels are expected, careful control over equipment parameters that define KPI metrics can effectively help optimize energy consumption levels in grinding circuits designed for specific product size standards.

In grinding circuits, over grinding of fine particles is a crucial cause for inefficient use of energy in comminution, as water and feed levels need to be carefully monitored since overly watered slurry causes fine particles to bypass the cyclones back into the mill and an overly dense slurry results in deficient slurry flow.

A tree-based machine learning technique was used to systematically evaluate and monitor the behaviour of key equipment parameters alongside their impact on the grinding. After careful tuning of the tree-based algorithms hyperparameters, the model’s ability to project and explain the behavior of grinding circuit equipment and its variability improves, yielding highly accurate results.

Further implementation of other machine learning techniques such as reinforcement learning and Deep Q Networks are developed in conjunction with the previous approach calibrated together towards lowering energy consumption and greenhouse gas emissions while preserving product quality and production volumes. The reinforcement learning framework trains a machine learning model to make a sequence of decisions, which aim to achieve a learning goal in uncertain and complex environments. Rewards and penalties are used to train this model in an environment where the learning goal is the reduction of greenhouse gas emissions from energy consumption, and the decisions are adjustments to be performed on key equipment parameters which are previously modelled by the tree-based algorithms. The changes made to the equipment parameters based on these decisions help maximize the efficient use of energy in the circuit.

The performance of the reinforcement learning algorithm directly depends on the reward function formulation as it directly influences the decision sequences. Combined with reinforcement learning, Deep Q Networks can be used to predict the rewards of each possible state change in the circuit. Deployed together, these machine learning techniques constitute a system that captures the knowledge of the grinding process by learning the trends of the effects of the variations in key equipment setting parameters over energy consumption.

After successfully training on historical data, the trained reinforcement learning model can suggest potential actions to modify key equipment settings that result in optimized energy consumption controlling the KPI metrics and ensuring product quality standards and production volumes.